1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c2-sse.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <smmintrin.h>
13
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/unaligned.h>
17
18
19
xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld64(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld64(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const uint8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 uint8_t* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(uint8_t) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 kc = round_up_po2(kc, 2 * sizeof(uint8_t));
42 const uint8_t* a0 = a;
43 uint8_t* c0 = c;
44
45 do {
46 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
47 w = (const void*) ((const int32_t*) w + 4);
48
49 size_t k = kc;
50 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
51 while (k >= 8 * sizeof(uint8_t)) {
52 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
53 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
54 a0 += 8;
55
56 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
57 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
58
59 vacc0x0123 = _mm_add_epi32(vacc0x0123,
60 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
61 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
62 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
63
64 vacc0x0123 = _mm_add_epi32(vacc0x0123,
65 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
66 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
67 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
68
69 vacc0x0123 = _mm_add_epi32(vacc0x0123,
70 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
71 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
72 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
73
74 vacc0x0123 = _mm_add_epi32(vacc0x0123,
75 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
76
77 w = (const void*) ((const uint8_t*) w + 32);
78 k -= 8 * sizeof(uint8_t);
79 }
80 if (k != 0) {
81 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
82 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
83 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
84
85 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
86 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
87 w = (const void*) ((const uint8_t*) w + 8);
88
89 vacc0x0123 = _mm_add_epi32(vacc0x0123,
90 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
91
92 if (k > 2 * sizeof(uint8_t)) {
93 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
94 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
95 w = (const void*) ((const uint8_t*) w + 8);
96
97 vacc0x0123 = _mm_add_epi32(vacc0x0123,
98 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
99
100 if (k > 4 * sizeof(uint8_t)) {
101 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
102 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
103 w = (const void*) ((const uint8_t*) w + 8);
104
105 vacc0x0123 = _mm_add_epi32(vacc0x0123,
106 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
107 }
108 }
109 }
110
111 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
112
113 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
114 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
115
116 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
117 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
118
119 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
120
121 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
122 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
123
124 __m128i vout = _mm_packus_epi16(vacc00x0123, vacc00x0123);
125
126 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
127
128 if (nc >= 4) {
129 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
130
131 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
132
133 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
134
135 nc -= 4;
136 } else {
137 if (nc & 2) {
138 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
139 c0 += 2;
140 vout = _mm_srli_epi32(vout, 16);
141 }
142 if (nc & 1) {
143 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
144 }
145
146 nc = 0;
147 }
148 } while (nc != 0);
149 }
150